How to generate DZYX_npy_f16 from the released npy_DZYX_complex radar tensors?

[bbljy]

Dear authors,

Thank you for releasing the RT-Pose dataset and code. I am currently
reproducing the HRRadarPose baseline using:

configs/cruw_pose/hr3d_one_hm_doppler.py

I found a difference between the released radar preprocessing output and
the radar data expected by the training code.

The released script data_processing/4Dradar2xyz.py saves the Cartesian
radar tensor to:

radar/npy_DZYX_complex

The saved tensor appears to be a complex-valued tensor with shape:

[64, 32, 128, 256] = [D, Z, Y, X]

However, the dataset loader used for training reads from:

DZYX_npy_f16

and directly converts the loaded array with:

arr_cube = np.load(...).astype(np.float32)

The main training configuration also uses:

RDR_TYPE = 'dzyx_real'
DZYX.NORMALIZING_VALUE = (0.0, 10.0)
backbone_cfg = 'hr_tiny_feat32_zyx_l4_in32'

Therefore, I would like to ask how the released 64-channel complex DZYX
tensor was converted into the DZYX_npy_f16 tensor used for training.

Could you please clarify the following details?

What is the exact shape and dtype of each DZYX_npy_f16 file?
How is the complex tensor converted to a real-valued tensor?
magnitude: abs(x)
power: abs(x) ** 2
real component
magnitude and phase
real and imaginary components
another representation
Was logarithmic or dB compression applied, such as:
log10(abs(x) + 1)
20 * log10(abs(x) + eps)
How were the 64 Doppler bins converted to the 32 input channels required
by hr_tiny_feat32_zyx_l4_in32?
selecting 32 bins
averaging adjacent bins
max pooling
positive/negative Doppler combination
another method
Was any background suppression, clipping, calibration, or dataset-level
normalization applied before saving the files?
Does NORMALIZING_VALUE=(0.0, 10.0) assume that the saved tensor has
already been transformed to approximately the range [0, 10]?
How is DZYX_npy_f16_complex generated for the phase/complex
configuration, and what tensor shape is expected by that configuration?

I have currently tested a custom preprocessing pipeline:

complex DZYX
-> magnitude
-> log10(x + 1)
-> retain all 64 Doppler channels
-> 64-channel HRNet3D

Using six sequences and training for 100 epochs, I obtained approximately:

MPJPE: 135.98 mm
ABS-MPJPE: 218.96 mm

However, I understand that this custom preprocessing may be different from
the preprocessing used in the paper.

Would it be possible to release the preprocessing script that generates
DZYX_npy_f16, or provide one corresponding pair of
npy_DZYX_complex and DZYX_npy_f16 sample files?

This missing conversion step appears to be important for reproducing the
reported result accurately.

Thank you very much for your help.